How it works

Every completed AFL match since 2015 is pulled from the official Champion Data StatsPro feed — one row per player per game with the full stat line (disposals, kicks, handballs, marks, tackles, goals, behinds, clearances, hit-outs, fantasy points) plus context: position, time-on-ground, venue, home/away and opponent.

For each game we build leakage-safe inputs — only what was known beforehand: rolling form (last 3/5/10 games), season- and career-to-date rates, role, a primary-ruck flag, rest days, and how much the opponent typically concedes. A game never sees its own result.

A gradient-boosting model predicts each player’s expected output for every market; goals and behinds use a Poisson formulation suited to low counts. Out of sample (seasons held back), the models beat a “last-5-games average” baseline on 9 of 10 markets. Fantasy points are modelled two ways — directly, and rebuilt from the components — as a cross-check.

Each match is simulated thousands of times. A team’s total for every stat is drawn around its expectation, then handed out to players by their share of the team — jittered each sim by the historical spread, so star and role players get realistic variance. Disposals are always kicks + handballs; fantasy is rebuilt from the simulated components. Team scores (6 × goals + behinds) give the win probability and totals. This yields a full, internally consistent distribution for every player and market.

The Compare page pulls live player-line odds from Sportsbet, TAB andLadbrokes and lays every book’s price side by side for each line, with the best highlighted. Next to them sits the model’s own probability and fair price, so you can see at a glance where a book is paying over the odds. (Dabble is a Pick’em product, not fixed odds, so it lives on the Pick’em page.)

From the best available price we report the edge (model probability − the market’s de-vigged probability) and expected value per dollar. The recommended stake uses the Kelly criterion — full Kelly maximises long-run growth; we default to a safer quarter-Kelly, sized to your bankroll and capped at 5% per bet. The Value page does the same for any price you enter by hand.

Models are imperfect: lineups change late, roles shift, and weather and tactics aren’t fully captured. Treat projections as one input, not certainty. Backtested accuracy and calibration are tracked so you can see where the model is reliable — and where it isn’t.